Machine-learnt state estimation for optimization in low voltage distribution grids

Real-time state estimation becomes essential for advanced monitoring and control of low-voltage networks. Machine-learning algorithms are fast predictors and have been employed in state estimation, but absence of adequate training datasets impede deployment. We present a novel approach for distribution system state estimation using machine learning methods, based on topological information as the sole input, no measurement data is required. The predictability of unknown quantities in low observability grid settings was studied in two test grids and one real-world example in various scenarios of observed and unobserved grid positions and loading configurations. The findings imply that accurate state estimation is possible even in low observability systems. The method allows to assess grids for the best number of measurement devices and their ideal positioning for maximum information gain.